Natively tackles local Whisper crashes and downloads that keep starting over
An October 15 source update changes how local ONNX workers stop and how interrupted model downloads resume. The maintainer reports macOS testing, with Windows still untested.
A stop request should not close the whole app
Natively has committed a local-speech reliability patch aimed at two frustrating failures: an ONNX worker stopping during a native call and aborting the desktop app, and a large model download losing its progress when the connection drops.
The October 15 commit links the shutdown work to a v2.9.2 crash report involving Whisper transcription. Instead of forcibly terminating a worker, the revised implementation sends a shutdown message that the worker handles when it is idle between native calls. The same approach covers local speech, embeddings, reranking and routing workers.
There is a tradeoff: the code deliberately avoids a hard-kill fallback. A worker stuck indefinitely in native code can remain alive until the app quits. The change targets a specific crash mechanism; it is not a claim that every local-model failure has disappeared.
Keep the bytes already downloaded
The new downloader writes large model files to a temporary partial file, requests the missing bytes with HTTP Range, and promotes a download only after its size matches the server’s stated size. The commit also describes download markers, damaged-model recovery, system-proxy support and a move to different exports of the same Whisper weights.
These are changes to the desktop harness and model-loading workflow, not a newly trained speech model. The maintainer reports testing interrupted and resumed downloads, Indonesian transcription and the Settings update row on macOS. The report explicitly says Windows was not executed. AITrending reviewed the original code and notes but did not reproduce those tests.
Source patch, existing installer
Natively documents macOS and Windows installation, including a Homebrew route on macOS. Its latest stable release is still the October 5 v2.9.2 package; the reviewed commit does not establish that downloaded installers already contain this patch.
Local transcription can avoid a cloud speech API key, while cloud-generated answers still depend on the configured model provider and its charges. The project also documents an Ollama-based offline option. Its source license permits personal and non-commercial uses, with separate permission required for commercial use; public source access is not a permissive open-source license.